Calibrated multi-view graph learning framework for infant cognitive abilities prediction

Abstract

Early prediction of cognitive development holds significant importance in neonatal healthcare, especially given the high incidence of cognitive deficits or developmental delays in preterm infants. Previous advances have already investigated the interior relation between brain cortical morphology and cognitive skills, leveraging this connection for prognostication. However, the small proportion of subjects with cognitive deficits in the cohort limits the predictive power of previous models, i.e., the data imbalance issue. To tackle this challenge, in this paper, we present the Calibrated Multi-view Graph Learning (CMGL) framework for cognition score prediction, a cortical graph learning model with capabilities for the imbalanced regression scenario. In order to collaboratively capture the morphological relations among brain regions, a multi-view cortical graph is constructed based on cortex developmental correlation and adaptive morphology similarity. On top of this graph, we train a diffusion graph convolutional backbone to obtain the cortical graph representation. Considering the data imbalance challenge, we propose a feature clustering module to calibrate the learned feature space, reducing training bias towards dominant classes. Moreover, we introduce smoothed reweighted mean absolute error loss based on label distribution smoothing to guide the training process in continuous imbalanced scenario. In the cross-validation experiment on our in-house dataset, the proposed CMGL achieves a mean square error of 0.1596, demonstrating state-of-the-art performance compared to other related methods.